Optical remote sensing image authenticity identification method and system based on multi-color space fusion

Abstract

The invention provides an optical remote sensing image authenticity identification method and system based on multi-color space fusion. The system comprises an initialization module, a multi-color space conversion module, a multi-model training module, a weight dynamic optimization module and a weighted evidence fusion module. An optical remote sensing image is input into an initialization module, and a color space is converted through a multi-color space conversion module; the multi-model training module is used for training a CNN model and outputting an authenticity probability result; the weight dynamic optimization module dynamically optimizes the weight of the CNN model of the multi-color space; and the weighted evidence fusion module fuses the authenticity probability result and outputs a final authenticity identification result. According to the method, the problem of missed detection of tampering traces caused by insufficient single color space information in a traditional method is solved, and the detection accuracy and robustness in a complex counterfeit scene are remarkably improved; interference of noise or low-quality characteristics is reduced, and high robustness is still kept under sensor difference or environment change; conflict evidences are dynamically adjusted, and decision reliability is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of optical remote sensing image authenticity identification, and particularly relates to an optical remote sensing image authenticity identification method and system based on multi-color space fusion. BACKGROUND

[0002] Remote sensing image authenticity identification is an important research direction in the field of remote sensing image processing, aiming to determine the authenticity of an image by analyzing its content and features. With the wide application of remote sensing technology in military, environmental, urban planning and other fields, its high value has given rise to forgery methods such as splicing, copying and pasting, and these tampering behaviors can mislead decision-making and threaten social security, so it is of great significance to develop effective identification technology.

[0003] Early methods mainly rely on manually designed features such as texture and edge, which can detect simple tampering but have limited effect on complex forgery (such as images generated by deep learning) and lack generalization ability. In recent years, methods based on convolutional neural networks (CNN) have significantly improved detection accuracy by automatically learning features, but existing researches are mostly limited to the RGB color space and fail to fully utilize the complementarity of different color spaces, limiting the further improvement of detection performance. For example, the HSV space is more sensitive to changes in hue and saturation, and may be more likely to capture tampering traces such as inconsistent lighting.

[0004] Current multi-feature fusion methods still have deficiencies in remote sensing authenticity identification: traditional weighted average or voting strategies cannot effectively balance the reliability of feature sources and are easily disturbed by noise; although evidence theory can model multi-source information through basic belief assignment (BBA) and achieve efficient fusion using the Dempster rule, its application in decision-level fusion of remote sensing image multi-color spaces has not been fully explored.

[0005] For example, the patent document "Joint Supervised Underwater Image Enhancement Algorithm Based on Detail Enhancement and Multi-Color Space Learning" (CN118982468A) discloses a joint supervised underwater image enhancement algorithm based on detail enhancement and multi-color space learning, which uses a structure and detail separation module, a T component extraction module, a multi-color feature extraction and fusion module, combines supervised and unsupervised learning methods, and uses an underwater imaging model for image enhancement to reduce color deviation and fogging phenomena. This method is susceptible to noise or low-quality features, and has poor robustness under sensor differences or environmental changes.

[0006] Therefore, a multi-color space evidence fusion method is proposed, which can reduce noise interference and significantly improve the accuracy and robustness of complex tampering detection. SUMMARY

[0007] Aiming at the defects in the prior art, the purpose of the present application is to provide an optical remote sensing image authenticity identification method and system based on multi-color space fusion.

[0008] According to the present application, an optical remote sensing image authenticity identification method based on multi-color space fusion is provided, comprising:

[0009] Step one, input the optical remote sensing image into the initialization module, and convert the color space through the multi-color space conversion module;

[0010] Step two, set parameters, train the CNN model through the multi-model training module, and generate authenticity probability results;

[0011] Step three, dynamically optimize the weight of the CNN model of the multi-color space through the weight dynamic optimization module;

[0012] Step four, fuse the authenticity probability results output by the CNN model according to the weight through the weighted evidence fusion module, and obtain the final authenticity identification result.

[0013] Preferably, the step one comprises:

[0014] Step S1.1, initialize the input RGB format optical remote sensing image

[0015] Step S1.2, convert the optical remote sensing image in the RGB color space into images in the HSV, YIQ and XYZ color spaces through the multi-color space conversion module.

[0016] In the step two, set the color space conversion parameters and the training hyperparameters of the CNN model, and train the CNN models with the same structure respectively using the optical remote sensing image in the RGB color space and the converted images in the HSV, YIQ and XYZ color spaces Extract complementary features and output authenticity probability results

[0017] The N training data are composed of images and corresponding authenticity labels.

[0018] Wherein, x p ∈R H×W×3 , represents the image in the RGB color space;

[0019] y p ∈{0,1}, represents the authenticity label, 0 is a real image, and 1 is a fake image;

[0020] H and W represent the width and height of the image respectively;

[0021] m p,n represents the image x pFor real and fake probabilities, for the vector.

[0022] Preferably, the step three comprises:

[0023] Step S3.1, initializing the weight;

[0024] Step S3.2, training the CNN model of RGB, HSV, YIQ and XYZ color space respectively, outputting the soft classification result and calculating the accuracy rate η n ;

[0025] Step S3.3, calculating the weight and optimizing the target according to the accuracy rate through a nonlinear mapping function;

[0026] Step S3.4, solving the optimal

[0027] The step four comprises:

[0028] Step S4.1, obtaining the soft classification result of the CNN model of RGB, HSV, YIQ and XYZ color space;

[0029] Step S4.2, performing BBA discount processing on the soft classification result;

[0030] Step S4.3, weighting DS evidence fusion on the soft classification result;

[0031] Step S4.4, converting the fused result into BetP probability form and making a decision.

[0032] Wherein, n represents the model serial number, n = 1, 2, 3, 4.

[0033] Preferably, the initialized weight β n ∈ [0, 1].

[0034] In the step S3.2, for the nth color space, the image is divided into training set and test set

[0035] Training the CNN model M n , the output soft classification result is The probability distribution of single-element focus element fake class ω1 and real class ω0 satisfies

[0036] The accuracy rate η n is calculated as:

[0037]

[0038] Wherein, Indicates the ith remote sensing training image in the training set;

[0039] This represents the j-th remote sensing test image in the test set;

[0040] These represent the true and false labels for the training and test images, respectively.

[0041] N tr N te These represent the number of images in the training set and the test set, respectively.

[0042] TP n TN n These represent the number of fake images and real images correctly classified by the model on the training set, respectively.

[0043] N tr This represents the total number of training data.

[0044] In step S3.3, the weight β n Through nonlinear mapping function The calculation and optimization objective is to minimize the mean squared error between the weighted DS fusion result and the true label:

[0045]

[0046] Where λ≥0 represents the penalty coefficient;

[0047] This represents the learned optimal λ value;

[0048] The one-hot encoded vector representing the actual label;

[0049] This represents the output of the weighted CNN model.

[0050] This represents the evidence synthesis operator based on the DS rule.

[0051] In step S3.4, the gradient descent method is used for iteration until the mean squared error loss converges or the set maximum number of iterations is reached.

[0052] The optimal To make η n Larger CNN models achieve relatively larger β. n And η n Smaller CNN models achieve relatively smaller β n of

[0053] Preferably, in step S4.1, four sets of test images in different color spaces are used. The inputs are fed into the corresponding CNN models respectively. In particular, we have The probabilities of being fake {ω1} and real {ω0} are

[0054] The Bayesian basic belief assignment (BBA) is

[0055]

[0056] where, denotes the jth remote sensing test image in the test set;

[0057] denotes the soft classification result of the test set;

[0058] N te denotes the total number of the test set;

[0059] denotes The probability of being fake in the corresponding color space;

[0060] denotes The probability of being real in the corresponding color space;

[0061] denotes a subset of the set Ω;

[0062] Ω denotes the global uncertainty set.

[0063] The step S4.2 according to the optimal weight β n The soft classification result (BBA) is discounted and simplified as

[0064]

[0065] where, denotes the BBA assigned to the subset A;

[0066] denotes the BBA assigned to the global uncertainty set Ω.

[0067] The rule of the weighted DS evidence fusion is

[0068]

[0069] The fusion result of the four groups of BBA is is:

[0070]

[0071] where, K denotes the conflict factor;

[0072] denotes a DS rule-based evidence combination operator;

[0073] B, C denote subsets of Ω;

[0074] m1 denotes the BBA of information source 1;

[0075] m2 denotes the BBA of information source 2;

[0076] denotes the discounted fusion result.

[0077] The BetP probability form of the fused result is:

[0078]

[0079] query image The authenticity decision is determined by comparing the size of the fused fake probability and the fused real probability The authenticity identification result is:

[0080]

[0081] Output the final authenticity identification result.

[0082] The application provides an optical remote sensing image authenticity identification system based on multi-color space fusion, comprising an initialization module, a multi-color space conversion module, a multi-model training module, a weight dynamic optimization module and a weighted evidence fusion module.

[0083] The optical remote sensing image is input into the initialization module, and the color space is converted by the multi-color space conversion module.

[0084] The multi-model training module trains the CNN model according to the set parameters to generate the authenticity probability result.

[0085] The weight dynamic optimization module dynamically optimizes the weight of the CNN model of the multi-color space.

[0086] The weighted evidence fusion module fuses the authenticity probability result output by the CNN model according to the weight to obtain the final authenticity identification result.

[0087] Preferably, the initialization module initializes the input RGB format optical remote sensing image The color space conversion parameters and the training hyperparameters of the CNN model are set.

[0088] The multi-color space conversion module converts the optical remote sensing image in the RGB color space into images in the HSV, YIQ and XYZ color spaces, respectively.

[0089] The multi-model training module trains the same structure of CNN model by using optical remote sensing images of RGB color space and converted images of HSV, YIQ and XYZ color spaces respectively The complementary features are extracted, and a true or false probability result is output

[0090] The N training data are composed of images and corresponding true or false labels of the images.

[0091] Wherein, x p ∈R H×W×3 , represents the image of RGB color space;

[0092] y p ∈{0, 1}, represents the true or false label, 0 is a real image, and 1 is a fake image;

[0093] H and W represent the width and height of the image respectively;

[0094] m p,n represents the image x p in the nth model, and p is the probability of being real and fake, which is a vector.

[0095] Preferably, the weight dynamic optimization module initializes the weight;

[0096] The CNN models of RGB, HSV, YIQ and XYZ color spaces are trained respectively, and soft classification results are output and the accuracy η n is calculated;

[0097] The weight is calculated and the target is optimized by a nonlinear mapping function according to the accuracy;

[0098] The gradient descent method is used to iteratively solve the optimal

[0099] The weighted evidence fusion module obtains the soft classification results of the CNN models of RGB, HSV, YIQ and XYZ color spaces;

[0100] The soft classification results are processed by BBA discounting;

[0101] The soft classification results are weighted and DS evidence fusion is performed;

[0102] The fused results are converted into BetP probability form and decision is made.

[0103] Wherein, n represents the model serial number, n = 1, 2, 3, 4.

[0104] Preferably, the initialized weight β n ∈[0, 1].

[0105] The multi-model training module divides the images into a training set and a test set for the nth color space and a test set

[0106] The CNN model M is trained n , and the output soft classification result is The probability distribution of the single-element focus element fake class ω1 and the real class ω0 satisfies

[0107] The weight dynamic optimization module calculates the accuracy η n as follows:

[0108]

[0109] wherein, represents the ith remote sensing training image in the training set;

[0110] represents the jth remote sensing test image in the test set;

[0111] and respectively represent the true and false labels of the training images and test images;

[0112] N tr , N te respectively represent the number of images in the training set and the test set;

[0113] TP n , TN n respectively represent the number of fake images and real images correctly classified by the model on the training set;

[0114] N tr represents the total number of training sets.

[0115] The weight β n is calculated by a nonlinear mapping function , and the optimization objective is to minimize the mean square error between the weighted DS fusion result and the true label:

[0116]

[0117] wherein, λ≥0, represents a penalty coefficient;

[0118] represents the optimal λ value learned;

[0119] represents the one-hot encoding vector of the true label;

[0120] represents the output result of the weighted CNN model;

[0121] represents the evidence synthesis operator based on the DS rule.

[0122] The gradient descent method is adopted for iteration until the mean square error loss converges or the maximum number of iterations is reached;

[0123] The optimal In order to make η n The relatively larger CNN model obtains the relatively larger β n And η n The relatively smaller CNN model obtains the relatively smaller β n The

[0124] Preferably, the weighted evidence fusion module inputs four groups of test images in different color spaces into the corresponding CNN model respectively, to obtain The probabilities of being a fake class {ω1} and a real class {ω0} are respectively

[0125] The Bayesian basic belief assignment BBA is:

[0126]

[0127] Wherein, represents the jth remote sensing test image in the test set;

[0128] represents the soft classification result of the test set;

[0129] N te represents the total number of test sets;

[0130] represents The probability of being a fake in the corresponding color space;

[0131] represents The probability of being real in the corresponding color space;

[0132] represents a subset of the set Ω;

[0133] Ω represents the global uncertainty set.

[0134] The weighted evidence fusion module discounts the soft classification result (BBA) according to the optimal weight β n And simplifies it to:

[0135]

[0136] Wherein, denotes the BBA assigned to A subset;

[0137] denotes the BBA assigned to global uncertainty set Ω.

[0138] The rule of the weighted DS evidence fusion is:

[0139]

[0140] The fusion result of the four groups of BBA is obtained is:

[0141]

[0142] Wherein, K represents a conflict factor;

[0143] denotes the evidence synthesis operator based on the DS rule;

[0144] B and C both denote subsets of set Ω.

[0145] m1 denotes the BBA of information source 1;

[0146] m2 denotes the BBA of information source 2;

[0147] denotes the discounted fusion result.

[0148] The BetP probability form of the fused result is:

[0149]

[0150] Query image The authenticity decision is determined by comparing the fused fake probability With the fused real probability The size of the authenticity identification result is:

[0151]

[0152] Output the final authenticity identification result.

[0153] Compared with the prior art, the present application has the beneficial effects as follows:

[0154] 1、The present application effectively solves the tampering trace missing detection problem caused by insufficient information of single color space in traditional methods through the complementary feature extraction of multi-color space and the evidence theory fusion mechanism, and significantly improves the detection accuracy and robustness in complex forgery scenes.

[0155] 2、The adaptive weight optimization of the application is based on the evidence theory to dynamically learn the reliability weight of each color space, reduces the interference of noise or low-quality features, and still maintains high robustness under sensor differences or environmental changes.

[0156] 3、The evidence conflict suppression of the application is through BBA discount processing to distribute the confidence of the low-weight model to the global uncertainty set, combines the DS rule normalization factor, dynamically adjusts the conflicting evidence, and improves the decision reliability. BRIEF DESCRIPTION OF DRAWINGS

[0157] Other features, objects, and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0158] Figure 1 A process schematic diagram of an optical remote sensing image authenticity identification method based on multi-color space fusion;

[0159] Figure 2 An optical remote sensing image authenticity identification schematic diagram based on multi-color space fusion;

[0160] Figure 3 A data set schematic diagram after color space conversion;

[0161] Figure 4 An optical remote sensing image authenticity identification system schematic diagram based on multi-color space fusion. DETAILED DESCRIPTION

[0162] The application will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the application, but do not limit the application in any form. It should be noted that, for those skilled in the art, without departing from the concept of the application, a number of changes and improvements can be made. These all belong to the protection scope of the application.

[0163] According to the application, an optical remote sensing image authenticity identification method based on multi-color space fusion is provided. Different tampering traces are strengthened by converting the image to HSV, YIQ, etc. color space, and a plurality of CNN models are trained, complementary features are extracted, and evidence theory is used for fusion, while optimizing the weight of each space to minimize the classification error. A new framework of multi-color space fusion is constructed, and decision-level information optimization is realized through evidence theory, solving the problem of insufficient single RGB space feature expression, and significantly improving the detection ability of deep learning generated forgery, complex splicing and other tampering means.

[0164] Specifically, taking the method for example, it includes four steps of multi-color space conversion, multi-model training, weight dynamic optimization and weighted evidence fusion. Figure 1

[0165] ​Step one, collect the optical remote sensing image with true and false label in RGB format, input the initialization module, and convert the RGB format image to HSV, YIQ and XYZ color space through the multi-color space conversion module.

[0166] Specifically, the source domain is initialized, and the input RGB image is processed and converted into images in HSV, YIQ and XYZ color space to extract complementary features in multi-color space, including:

[0167] Step S1.1, initialize the input image, and the input optical remote sensing image dataset is an RGB format image.

[0168] RGB color space, as the most popular color space, has three channels (R, G, B), usually represented by 24 bits, each channel represented by 8 bits, with a value range of 0 to 255. In the RGB color space, all colors can be represented by different shades of red, green and blue. The input optical remote sensing image dataset is Wherein, x p ∈R H×W×3 , represents the image in RGB color space, H and W represent the width and height of the image respectively. y p ∈{0,1}, represents the true and false label, 0 is a real image and 1 is a fake image.

[0169] Step S1.2, color space domain conversion, for example, convert the RGB color space picture to the HSV color space picture, or in other words, RGB to CIEHSV color space conversion. Figure 3

[0170] In practical applications, considering human perception of color, hue-saturation-value (HSV) color space is developed to describe color through saturation and lightness. The color space conversion parameters include hue (H), saturation (S), lightness (V) channel range, YIQ luminance (Y), in-phase (I), and orthogonal (Q) channel matrix, XYZ CIE1931 standard conversion matrix, and definition of CNN model structure (such as ResNet-50) and training hyperparameters.

[0171] In more preferred examples, the learning rate α = 0.01, the batch size B = 64, and the number of iterations T = 100 in the training hyperparameters.

[0172] The color space domain conversion includes:

[0173] Normalize the RGB value, normalize the RGB channel value from [0-255] to [0,1] interval:

[0174] R′=R / 255 ​

[0175] G' = G / 255

[0176] B' = B / 255

[0177] Where R', G', B' are the normalized values of R, G, B respectively.

[0178] Compute the intermediate variables, C max = max{R', G', B'}, C min = min{R', G', B'}, Δ = C max - C min .

[0179] Where: C max represents the maximum value in the RGB channel, which determines the lightness (V). C min represents the minimum value in the RGB channel, which is used to calculate the saturation. Δ represents the span of the RGB channel value, which is used to determine the hue.

[0180] According to the results of the intermediate variables, calculate the hue (H), saturation (S), lightness (V).

[0181] Hue (H) calculation: when Δ = 0, H is 0°, indicating that the gray color has no hue, and different C max correspond to different hue calculation formulas, ensuring that H is in the range of 0°-360°.

[0182]

[0183] Saturation (S) calculation: S represents the purity of the color, the larger the value, the more vibrant the color, ranging between 0%-100%.

[0184]

[0185] Lightness (V) calculation: V represents the brightness, determined by C max , ranging between 0%-100%.

[0186] V = C max

[0187] Take pixel R = 128, G = 64, B = 32 as an example, first normalize to get R' = 0.502, G' = 0.251, B' = 0.125. Calculate the intermediate variables to get C max = 0.502, C min = 0.125, Δ = 0.377. Finally, calculate the values of H, S, V according to these intermediate variables, since C max = R', therefore Since C max≠ 0, so S = 0.377 / 0.502 ~ 0.751 (i.e. 75.1%); V = C max = 0.502.

[0188] Step S1.3, convert the image in RGB color space to the image in YIQ (CIE 1931 YIQ) color space.

[0189] YIQ is a color space commonly used in television broadcasting, in which Y represents the brightness or lightness of the image, and I and Q are the chroma components. I reflects the in-phase component, which approximately represents the amount of blue or orange hue in the image, and Q represents the quadrature component, which approximately represents the amount of green or purple hue in the image. The pixels of the image in YIQ color space are obtained from the RGB color space in the following way:

[0190]

[0191] where Y represents the brightness component, which is obtained by weighted average of RGB, and the weights are based on the sensitivity of the human eye to different colors. The larger the value of Y, the higher the brightness of the image.

[0192] I represents the blue-orange chroma component, ranging from [-0.596 x 255, 0.596 x 255], with negative numbers representing blue hue and positive numbers representing orange hue. As I increases, the saturation of the color also increases.

[0193] Q represents the green-purple chroma component, ranging from [-0.523 x 255, 0.523 x 255], with negative numbers representing green hue and positive numbers representing purple hue. As Q increases, the saturation of the color also increases.

[0194] Taking a pixel R = 128, G = 64, B = 32 as an example, the calculation results are:

[0195] Y = 0.299 x 128 + 0.587 x 64 + 0.114 x 32 ~ 78.3

[0196] I = 0.596 x 128 - 0.274 x 64 - 0.322 x 32 ~ 47.2

[0197] Q = 0.211 x 128 - 0.523 x 64 + 0.312 x 32 ~ -5.8

[0198] The color of the calculation result is relatively bright, the in-phase component is biased towards orange hue, and the quadrature component is biased towards green hue.

[0199] Step S1.4, convert the image in RGB color space to the image in XYZ (CIE 1931 XYZ) color space.

[0200] XYZ is a CIE-defined device-independent color space, commonly used in the field of color science research. The purpose of designing XYZ color space is to cover all colors visible to the human eye, while being device-independent. Since the red, green and blue primary color lights selected from the actual spectrum cannot be mixed to match all colors in nature, CIE proposes theoretical primary colors XYZ to match all colors in theory and represent colors with non-negative values. XYZ tristimulus values are obtained by linear transformation of RGB color space, which is equivalent to using XYZ basis to replace RGB basis to represent colors. The linear transformation method is as follows:

[0201]

[0202] Where Y represents a direct correlation with brightness, similar to the brightness in HSV / YIQ. X and Z both represent chroma components, which are used to avoid negative values and cover a wider color range. X represents the red-green component, which is related to the perception of long-wavelength light, and Z represents the blue component of color, which is related to the perception of short-wavelength light.

[0203] Taking the pixel R=128, G=64, B=32 as an example, the calculation results are as follows:

[0204] X=0.412*128+0.358*64+0.180*32≈78.5

[0205] Y=0.213*128+0.715*64+0.072*32≈78.3

[0206] Z=0.019*128+0.119*64+0.950*32≈36.2

[0207] The value of Y component in XYZ is the same as that of Y component in YIQ, and the color is relatively bright. It is dominated by red-green and is a warm color tone close to orange-yellow.

[0208] Step two, set the color space conversion parameters and the training hyperparameters of the CNN model. The multi-model training module uses these different color space images to train the same structure of convolutional neural network (CNN) model, and generates the confidence prediction results of each model for image authenticity.

[0209] Specifically, the converted HSV, YIQ and XYZ images are used to train CNN models, and complementary features of multiple color spaces are extracted. The training uses N training data, which consists of images and their corresponding authenticity labels.

[0210] The existing convolutional neural network (CNN) models are trained by the remote sensing images in RGB, HSV, YIQ and XYZ color spaces respectively. The images in different color spaces usually perform differently when trained by the CNN model with the same network structure. That is, the CNN model trained by the images in HSV, YIQ and XYZ color spaces can provide some complementary information for the CNN model trained by the images in RGB color space only. Some specific categories of query images can be more accurately classified by the CNN model trained by the images in HSV, YIQ or XYZ color space.

[0211] Therefore, the performance of authenticity identification will be improved by making full use of the four CNN models instead of using only the CNN model trained in RGB color space. Four sets of authenticity judgment results are obtained for a query image by adopting the CNN models trained in different color spaces. In this case, the authenticity judgment results are complementary to each other, and their combination can effectively integrate the advantages of different color space images, thereby improving the performance and accuracy of remote sensing image authenticity detection.

[0212] Training the CNN models with the same structure on the converted HSV, YIQ and XYZ images respectively (outputting authenticity probability results)

[0213] wherein m p,n is a vector, indicating the confidence of the image x p in the nth model.

[0214] For example, M = [0.2 0.8], which means the probability of being real is 0.2 and the probability of being fake is 0.8.

[0215] Step three, the evidence weight learning module, i.e. the weight dynamic optimization module, dynamically optimizes and adjusts the weights of the CNN models in different color spaces by minimizing the mean square error (MSE) between the fusion result and the labeled authenticity label.

[0216] In the multi-source information fusion process based on evidence theory, the reliability of the soft classification results (BBA) generated by the CNN models in different color spaces is significantly different. In order to ensure that the high-precision models occupy higher decision weights in the fusion process, the weights of each model need to be dynamically optimized to give different confidence. According to the classification performance of each CNN model on the labeled data set, the optimal weight parameter is automatically learned. Specifically, the following steps are included:

[0217] Step S3.1, initialize the weight. Assign an initial weight β n ∈ [0, 1] to each color space model, which satisfies to embody the importance difference of different color spaces.

[0218] where N represents the number of color spaces.

[0219] Step S3.2, model accuracy η n The definition and calculation. Model reliability is quantified by the accuracy η n of its classification on the labeled dataset. Let the training set and test set in the RGB color space be

[0220] where and represent the i-th and j-th remote sensing images in the training set and test set, respectively, and N tr , N te represent the number of images in the training set and test set, respectively. In the training set, is a training image, is the true and false label of the training image. In the test set, is a test image, and represent the true and false label of the test image.

[0221] Specifically, for the nth color space (including RGB, HSV, YIQ, XYZ), the soft classification result output by the trained CNN model M n is , which represents the confidence that image is judged as fake, and is regarded as the Bayesian basic belief assignment (BBA) in evidence theory, i.e., the probability distribution of single-element focal elements fake class {ω1} and real class {ω0}, which satisfies

[0222] To measure the reliability of each model, the classification accuracy η n of the nth color space CNN model on the training set is defined as:

[0223]

[0224] where TP n , TN n are the number of fake images and real images correctly classified by the model on the training set, and N tr is the total number of training sets. η n is calculated by cross-validation or hold-out method.

[0225] Step S3.3, nonlinear weight mapping function and optimization objective weight, weight β n is calculated by a nonlinear mapping function , where λ ≥ 0 is a penalty coefficient used to adjust the amplitude of the weight growth with the accuracy.

[0226] Specifically, the optimization objective is to minimize the mean square error (MSE) between the weighted DS fusion result and the true label:

[0227]

[0228] wherein, represents the encoding vector of the true label, represents the evidence synthesis operator based on the DS rule, represents the output result of the weighted CNN model.

[0229] Step S3.4, iteratively solve the optimal so that the high-accuracy model (η n is relatively large) obtains a higher weight (β n is relatively small). n

[0230] Forward propagation, calculate the weight under the current λ and perform weighted DS fusion on the training set samples to obtain the prediction probability

[0231] Calculate the MSE loss:

[0232]

[0233] Back propagation, calculate the gradient of the loss λ:

[0234]

[0235] Parameter update:

[0236]

[0237] wherein, α is the learning rate.

[0238] Repeat the above steps continuously, iterate until the loss converges or reaches the maximum number of iterations, and obtain the optimal

[0239] In more preferred examples, the accuracies of the four color space models are η1=0.85 (RGB), η2=0.78 (HSV), η3=0.72 (YIQ), and η4=0.65 (XYZ), and the learning rate α=0.01.

[0240] First iteration:

[0241] β n =[0.85, 0.78, 0.72, 0.65], the sum of the weights is 1, and after normalization:

[0242] ​β n = [0.298, 0.273, 0.253, 0.176]

[0243] Calculate the fusion result The loss is L = 0.15, and the gradient calculation is Update λ = 1.02.

[0244] Second iteration, calculate:

[0245] β n = [0.85 1.02 , 0.78 1.02 , 0.72 1.02 , 0.65 1.02 ] ≈ [0.848, 0.776, 0.715, 0.642]

[0246] The sum of the weights is 1, and after normalization:

[0247] β n = [0.300, 0.275, 0.253, 0.172]

[0248] Repeat the above steps, calculate the loss, update, and get the optimal

[0249] Step four, the weighted evidence fusion module based on evidence theory and learned weights on the multi-color space of the fake discrimination result, that is, the prediction result of the multi-model, is fused, and the final true and false identification result is output.

[0250] Through the dynamically optimized weight β n , the soft classification result of the test image of the multi-color space is weighted evidence fusion by evidence theory, and the final true and false identification result is generated.

[0251] Step S4.1, obtain the classification result. Four groups of images of the test image converted by each color space are input into the corresponding pre-trained CNN model to obtain

[0252] The probabilities of the fake class {ω1} and the real class {ω0} output by the model are respectively Wherein, represents the probability of being fake in the corresponding color space. represents the probability of being real in the corresponding color space. The BBA of each model is represented as:

[0253]

[0254] wherein, denotes a subset of the set Ω, Ω denotes the global uncertainty set.

[0255] Step S4.2, BBA discounting processing. Since the BBA of low reliability model often causes greater conflict in fusion, in order to prevent unreliable BBA from causing conflict in DS combination, the optimal weight β n The BBA is discounted and weighted. The specific formula is:

[0256]

[0257] Since the focal element of Bayesian BBA is only a single-element set, Ω itself is not allocated a confidence, so the discounted BBA is simplified as:

[0258]

[0259] At the same time, the remaining unallocated confidence 1-β n is allocated to the global uncertainty set Ω, that is:

[0260]

[0261] Step S4.3, weighted DS evidence fusion. The four groups of discounted BBAs are synthesized by using the Dempster-Shafer (DS) rule. For the synthesis of two BBAs, the DS rule is defined as:

[0262]

[0263] wherein, the conflict factor is:

[0264]

[0265] B, C both denote a subset of the set Ω, m1 denotes the basic belief assignment (BBA) of the information source 1, and m2 denotes the basic belief assignment (BBA) of the information source 2.

[0266] Since all the BBAs in the step are Bayesian BBAs, that is, only {ω0} and {ω1} are contained, the synthesis of the four groups of BBAs can be realized by two-by-two fusion. The four groups of discounted BBAs are synthesized by using the DS rule, and the final fusion result is:

[0267]

[0268] wherein, is the discounted fusion result.

[0269] Step S4.4, Pignistic (BetP) probability transformation and decision making. To generate the final classification probability, the synthesized BBA needs to be transformed into BetP probability form to support decision making. For the BetP probability of According to the definition of BetP:

[0270]

[0271] Since The focal elements are only {ω0}, {ω1} and Ω, so the BetP probability is simplified as:

[0272]

[0273] Where X represents the focal element, ω c represents the focal element with 1 element, c represents the index subscript, and C represents the number of categories.

[0274] Finally, the authenticity decision of the query image is determined by comparing and :

[0275]

[0276] In more preferred examples, the test image The CNN model outputs in RGB, HSV, YIQ, and XYZ color spaces are respectively:

[0277]

[0278] The optimal weight learned is β n = [0.35, 0.30, 0.25, 0.10].

[0279] Calculate the BBA discount processing, in RGB space:

[0280]

[0281] Calculate the BBA discount processing in other color spaces

[0282]

[0283] Weighted DS evidence fusion, first fusion

[0284] Conflict factor: K 1,2 = (0.2975 * 0.0900) + (0.0525 * 0.2100) = 0.0378

[0285] Synthesis result:

[0286]

[0287] Repeat the above fusion process, and the final fusion result is m j ({ω0}) = 0.1837, m j ({ω1}) = 0.4956. Perform BetP probability conversion:

[0288]

[0289] Since Therefore, the image is determined to be fake.

[0290] In more preferred examples, the color space is converted to Figure 2 For example, using the ResNet50 deep neural network architecture, on the Fake-LoveDA remote sensing image dataset, four independent classification models are constructed through end-to-end training using RGB, HSV, YIQ, and XYZ color space conversion strategies. m1, m2, m3, and m4 represent the BBA in the RGB, HSV, YIQ, and XYZ color spaces, respectively. The test results are shown in Table 1, and the model based on the original RGB space achieved an overall accuracy (OA) of 83.62%. The mixed retraining of the ResNet50 neural network model using images from the four color spaces achieved a mixed retraining overall accuracy (DC) of 84.56%. The weighted average of the soft outputs of the models trained on images from the four color spaces achieved an average overall accuracy (AF) of 82.87%, demonstrating the advantages of deep convolutional neural networks in extracting original color features.

[0291] Table 1, Fake-LoveDA remote sensing image dataset color space test results:

[0292]

[0293] Through the multi-model fusion strategy, an overall accuracy of 84.79% (ECMS) is finally achieved, which is 1.17 percentage points higher than the optimal single model, verifying the complementary value of multi-color space feature fusion. While maintaining the simplicity of the model, the feature diversity is enhanced through color space transformation, providing a new technical approach for remote sensing image authenticity identification.

[0294] The combination of deep features and multi-color space representation can effectively capture tampering traces, especially in some common forgery methods, and exhibit stable detection performance, which has important application value for improving the credibility of remote sensing data.

[0295] The application also provides an optical remote sensing image authenticity identification system based on multi-color space fusion.

[0296] The optical remote sensing image authenticity identification system based on multi-color space fusion provided by the application comprises Figure 4 For example, the system comprises an initialization module, a multi-color space conversion module, a multi-model training module, a weight dynamic optimization module and a weighted evidence fusion module.

[0297] The data is input into the initialization module and the parameters are configured, and then the multi-color space conversion, model training, weight optimization and evidence fusion are sequentially performed to finally realize accurate identification of the authenticity of the optical remote sensing image. Figure 4 In the formula, X r(n) represents the nth original RGB input image, m1, m2, m3 and m4 represent the BBA in the RGB, HSV, YIQ and XYZ color spaces respectively, X rn , X hn , X yn , X xn , respectively represent the converted images in the RGB, HSV, YIQ and XYZ color spaces.

[0298] The initialization module is used for inputting an optical remote sensing image dataset, including images in RGB format and corresponding authenticity labels, and setting color space conversion parameters and training hyperparameters of a CNN model.

[0299] Specifically, the initialization module initializes the input RGB format optical remote sensing image The color space conversion parameters and the training hyperparameters of the CNN model are set.

[0300] The multi-color space conversion module comprises an RGB-to-HSV conversion unit, an RGB-to-YIQ conversion unit and an RGB-to-XYZ conversion unit, which respectively convert the input RGB image into the HSV, YIQ and XYZ color spaces, and realize complementary feature extraction of the multi-color space by calculating hue, saturation and lightness components.

[0301] The multi-model training module trains the CNN models with the same structure by using the converted HSV, YIQ and XYZ images, respectively, to generate confidence prediction results of the models on the authenticity of the images.

[0302] Specifically, the multi-model training module trains CNN models with the same structure using RGB format optical remote sensing images and converted images in HSV, YIQ and XYZ color spaces respectively Extract complementary features and output authenticity probability results

[0303] For the nth color space, the image is divided into a training set and a test set

[0304] Train the CNN model M n , and the output soft classification result is The probability distribution of the single-element focal element fake class ω1 and the real class ω0 satisfies

[0305] The N training data are composed of images and corresponding authenticity labels of the images;

[0306] where x p ∈R H×W×3 , represents an image in the RGB color space;

[0307] y p ∈{0, 1}, represents an authenticity label, 0 for a real image and 1 for a fake image;

[0308] H and W represent the width and height of the image respectively;

[0309] m p,n represents the probability that the image x p is real and fake in the nth model, and is a vector.

[0310] The weight dynamic optimization module dynamically optimizes the weights of the CNN models in each color space through a weight initialization unit, a target function definition unit, a gradient calculation and update unit, and a convergence judgment unit. First, the module initializes the weights and constructs a mean square error loss function and a regularization term, then iteratively updates the weights through the gradient descent method, and finally outputs the optimal weight distribution.

[0311] Specifically, the weight dynamic optimization module initializes the weights;

[0312] The CNN models in RGB, HSV, YIQ and XYZ color spaces are trained respectively, and the soft classification results are output and the accuracy η n is calculated; wherein n represents the model number, n = 1, 2, 3, 4.

[0313] The weights are calculated and optimized according to the accuracy through a nonlinear mapping function; the gradient descent method is used to iteratively solve the optimal

[0314] In more preferred examples, the initialization weight β n ∈ [0, 1] satisfies

[0315] The weight dynamic optimization module calculates the accuracy rate η n as follows:

[0316]

[0317] wherein, represents the i-th remote sensing training image in the training set;

[0318] represents the j-th remote sensing test image in the test set;

[0319] respectively represent the true and false labels of the training images and test images;

[0320] N tr , N te respectively represent the number of images in the training set and test set;

[0321] TP n , TN n respectively represent the number of fake images and real images correctly classified by the model on the training set;

[0322] N tr represents the total number of the training set.

[0323] The weight β n is calculated by a nonlinear mapping function The optimization objective is to minimize the mean square error between the weighted DS fusion result and the true label:

[0324]

[0325] wherein, λ ≥ 0 represents a penalty coefficient;

[0326] represents the optimal λ value learned;

[0327] represents the one-hot encoding vector of the true label;

[0328] represents the output result of the weighted CNN model;

[0329] represents the evidence synthesis operator based on the DS rule.

[0330] Gradient descent method is adopted for iteration until the mean square error loss converges or the maximum number of iterations set is reached.

[0331] The optimal To make η n Larger CNN models achieve relatively larger β. n And η n Smaller CNN models achieve relatively smaller β n of

[0332] The weighted evidence fusion module includes a basic confidence assignment construction unit, a Dempster combination rule unit, and a decision unit. After converting the prediction results of each model into basic confidence assignments, it fuses them based on evidence theory and outputs the final authenticity identification result through probability transformation and threshold judgment.

[0333] Specifically, the weighted evidence fusion module obtains the soft classification results of the CNN model in the RGB, HSV, YIQ and XYZ color spaces;

[0334] Apply BBA discounting to the soft classification results;

[0335] Weighted DS evidence fusion of soft classification results;

[0336] The fused results are transformed into BetP probabilistic form for decision-making.

[0337] In more preferred embodiments, the weighted evidence fusion module combines four sets of test images from different color spaces. The inputs are fed into the corresponding CNN models respectively. In the middle, the obtained The probabilities of being a fake class {ω1} and a real class {ω0} are respectively

[0338] The Bayesian basic confidence assignment (BBA) is as follows:

[0339]

[0340] in, This represents the j-th remote sensing test image in the test set;

[0341] This indicates the soft classification result of the test set;

[0342] N te Indicates the total number of test sets;

[0343] express In the corresponding color space, this represents the probability of forgery;

[0344] express In the corresponding color space, this represents the true probability.

[0345] denotes a subset of the set Ω;

[0346] Ω denotes a global uncertainty set.

[0347] The weighted evidence fusion module fuses the BBA according to the optimal weight β n The soft classification result is discounted by BBA and simplified as:

[0348]

[0349] wherein, denotes the BBA assigned to the subset A;

[0350] Ω denotes the BBA assigned to the global uncertainty set Ω.

[0351] The rule of the weighted DS evidence fusion is:

[0352]

[0353] The four groups of BBA are fused two by two and synthesized, and the final fusion result is:

[0354]

[0355] wherein, K denotes a conflict factor;

[0356] denotes an evidence synthesis operator based on the DS rule;

[0357] B and C both denote a subset of the set Ω;

[0358] m1 denotes the BBA of the information source 1;

[0359] m2 denotes the BBA of the information source 2;

[0360] denotes the discounted fusion result.

[0361] The BetP probability form of the fused result is:

[0362]

[0363] Query the image for a true or false decision, and determine the true or false identification result by comparing the size of the fused fake probability and the fused real probability

[0364]

[0365] ​Output the final authenticity identification result.

[0366] In more preferred examples, adaptive threshold adjustment function is also supported to dynamically adjust the decision threshold according to the actual application scenario to balance the false rejection rate and the false acceptance rate. At the same time, it has high scalability and can flexibly access new color space or detection model, and is suitable for the authenticity identification needs of multi-source remote sensing data such as satellites and aircrafts.

[0367] Through the multi-color space complementary feature extraction and evidence theory fusion mechanism, the problem of missing tamper traces caused by insufficient information of a single color space in traditional methods is effectively solved, and the detection accuracy and robustness in complex fake scenes are significantly improved.

[0368] Those skilled in the art know that, in addition to implementing the system provided by the present application and each device, module and unit thereof in a pure computer readable program code manner, the system provided by the present application and each device, module and unit thereof can also be implemented in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers and embedded microcontrollers by logically programming the method steps to achieve the same functions. Therefore, the system provided by the present application and each device, module and unit thereof can be considered as a hardware component, and the devices, modules and units included therein for achieving various functions can also be considered as structures within the hardware component; the devices, modules and units for achieving various functions can also be considered as both software modules for implementing methods and structures within the hardware component.

[0369] The specific embodiments of the present application are described above. It needs to be understood that the present application is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essential content of the present application. In the case of no conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. An optical remote sensing image authenticity identification method based on multi-color space fusion, characterized in that, The method comprises the following steps: Step 1: inputting an optical remote sensing image into an initialization module, and converting a color space through a multi-color space conversion module; Step 2: setting parameters, training a CNN model through a multi-model training module, and generating a true or false probability result; Step 3: dynamically optimizing the weight of the CNN model of the multi-color space through a weight dynamic optimization module; Step 4: fusing the true or false probability result output by the CNN model according to the weight through a weighted evidence fusion module, and obtaining a final true or false identification result. 2.The method according to claim 1, wherein, The step 1 comprises: Step S1.1, initializing the input optical remote sensing image in RGB format Step S1.2: converting the optical remote sensing image in the RGB color space into images in the HSV, YIQ and XYZ color spaces respectively through the multi-color space conversion module; In the second step, color space conversion parameters and training hyperparameters of the CNN model are set, and the same structure of the CNN model is trained by using optical remote sensing images in the RGB color space and converted images in the HSV, YIQ and XYZ color spaces The complementary features are extracted, and a true or false probability result is output The N training data are composed of images and corresponding true or false labels of the images; where x p ∈ R H×W×3 denotes an image in the RGB color space; y p ∈ {0,1}, represents the authenticity label, 0 is a real image, 1 is a fake image; H and W respectively represent the width and height of the image; m p,n represents the image x in the nth model p is the probability of being real and fake, is a vector. 3.The method according to claim 1, wherein, The step 3 comprises: Step S3.1: initializing the weight; Step S3.2, training the CNN model of RGB, HSV, YIQ and XYZ color space respectively, outputting soft classification results and calculating accuracy rate η n ; Step S3.3: calculating the weight and optimizing the target through a nonlinear mapping function according to the accuracy rate; Step S3.4, iteratively solve the optimum by gradient descent method The step 4 comprises: Step S4.1: obtaining soft classification results of the CNN models in the RGB, HSV, YIQ and XYZ color spaces; Step S4.2: performing BBA discount processing on the soft classification results; Step S4.3: performing weighted DS evidence fusion on the soft classification results; Step S4.4: converting the fused results into BetP probability form and making a decision; Wherein, n represents the model serial number, n = 1, 2, 3, 4.

4. The method according to claim 3, wherein The initialization weight β n ∈ [0, 1]; In said step S3.2, for the n-th color space, the image is divided into a training set and a test set Training a CNN model M n The output soft classification result is The probability distributions of the single-element focus element fake class ω1 and the real class ω0 satisfy Computing the accuracy η n is: wherein, denotes the i-th remote sensing training image in the training set; represents the jth remote sensing test image in the test set; respectively represent the authenticity labels of the training images and the test images; N tr , N te respectively represent the number of images in the training set, test set; TP n , TN n respectively denote the number of fake images, real images correctly classified by the model on the training set; N tr denotes the total number of training sets; The weight β in the step S3.3 n By a non-linear mapping function The optimization objective is to minimize the mean square error between the weighted DS fusion result and the true label: Wherein, λ≥0, represents a penalty coefficient; represents the optimal value of λ learned; one-hot encoded vector representing the true label; represents the output result of the empowered CNN model; represents a DS rule-based evidence combination operator; In the step S3.4, the gradient descent method is used for iteration until the mean square error loss converges or the maximum number of iterations set is reached; The optimal To make η n The larger CNN model obtains a relatively larger β n While η n The smaller CNN model obtains a relatively smaller β n The 5. The method according to claim 3, wherein the method is characterized by, The four groups of test images of different color spaces in step S4.1 are respectively input into the corresponding CNN model , and the probabilities of the obtained forgery class {ω1} and real class {ω0} are respectively The Bayesian basic belief assignment BBA is: wherein, denotes the jth remote sensing test image in the test set; a soft classification result representing the test set; N te denotes the total number of test sets; representing in the corresponding color space, a probability of forgery; representing is the probability of being real in the corresponding color space; a subset of the set Ω; Ω represents a global uncertainty set. The step S4.2 according to the optimal weight β n The soft classification result is discounted and simplified as: wherein, represents the BBA assigned to the A subset; denotes the BBA assigned to the global uncertainty set Ω; The rule of the weighted DS evidence fusion is: Fusion of the four groups of BBA results in the fusion result is: Wherein, K represents a conflict factor; represents a DS rule-based evidence synthesis operator; B and C both represent subsets of the set Ω; m1 represents the BBA of the information source 1; m2 represents the BBA of the information source 2; represents the fused result after discounting; The BetP probability form of the fused results is: Query image The authenticity decision is made by comparing the fused probability of forgery with the fused probability of authenticity to determine the authenticity discrimination result: Output the final true or false identification result.

6. An optical remote sensing image authenticity identification system based on multi-color space fusion, characterized in that, The method comprises the following steps: An initialization module, a multi-color space conversion module, a multi-model training module, a weight dynamic optimization module and a weighted evidence fusion module are provided; An optical remote sensing image is input into the initialization module, and a color space is converted through the multi-color space conversion module; A multi-model training module trains a CNN model according to the set parameters, and generates a true or false probability result; A weight dynamic optimization module dynamically optimizes the weight of the CNN model of the multi-color space; A weighted evidence fusion module fuses the true or false probability result output by the CNN model according to the weight, and obtains a final true or false identification result.

7. The multi-color space fusion based optical remote sensing image authenticity discrimination system according to claim 6, characterized in that, The initialization module initializes an input optical remote sensing image in RGB format Set color space conversion parameters and training hyperparameters of the CNN model; The multi-color space conversion module converts the optical remote sensing image in the RGB color space into images in the HSV, YIQ and XYZ color spaces respectively; The multi-model training module respectively trains the CNN models with the same structure by using the optical remote sensing image in the RGB color space and the converted images in the HSV, YIQ and XYZ color spaces The complementary features are extracted, and a probability result of authenticity is output The N training data are composed of images and corresponding true or false labels of the images; where x p ∈R H×W×3 denotes an image in the RGB color space; y p ∈ {0,1}, represents the authenticity label, 0 is a real image, 1 is a fake image; H and W respectively represent the width and height of the image; m p,n denotes the image x in the nth model p is the probability of being real and fake, is a vector.

8. The multi-color space fusion based optical remote sensing image authenticity identification system according to claim 6, characterized in that, The weight dynamic optimization module initializes the weight; Train the CNN model of RGB, HSV, YIQ and XYZ color space respectively, output soft classification results and calculate accuracy rate η n ; The weight is calculated through a nonlinear mapping function according to the accuracy rate, and the target is optimized; The optimal solution is solved iteratively using the gradient descent method The weighted evidence fusion module obtains soft classification results of the CNN models in the RGB, HSV, YIQ and XYZ color spaces; The soft classification result is subjected to BBA discount processing; The soft classification result is subjected to weighted DS evidence fusion; The fused result is converted into BetP probability form and decision is made; Wherein, n represents model serial number, n=1, 2, 3, 4.

9. The multi-color space fusion based optical remote sensing image authenticity discrimination system according to claim 8, characterized in that, The initialization weight β n ∈ [0, 1] The multi-model training module divides the images into a training set and a test set for the nth color space and a test set Training a CNN model M n The output soft classification result is The probability distributions of the single-element focus element fake class ω1 and the real class ω0 satisfy The weight dynamic optimization module calculates the accuracy rate η n is: wherein, denotes the i-th remote sensing training image in the training set; represents the jth remote sensing test image in the test set; respectively represent the authenticity labels of the training images and the test images; N tr , N te respectively represent the number of images in the training set, test set; TP n , TN n respectively denote the number of fake images, real images correctly classified by the model on the training set; N tr denotes the total number of training sets; by a non-linear mapping function computing weights β n The optimization objective is to minimize the mean square error between the weighted DS fusion result and the ground truth label: Wherein, λ≥0, represents penalty coefficient; λoptdenotes the optimal λ value learned by the learning algorithm; one-hot encoded vector representing the true label; represents the output result of the empowered CNN model; represents a DS rule-based evidence synthesis operator; Gradient descent method is adopted for iteration until mean square error loss converges or the maximum iteration number set is reached; The optimal To make η n The larger CNN model obtains a relatively larger β n While η n The smaller CNN model obtains a relatively smaller β n The 10. The multi-color space fusion based optical remote sensing image authenticity discrimination system according to claim 8, characterized in that, The weighted evidence fusion module combines four sets of test images from different color spaces. The inputs are fed into the corresponding CNN models. In the middle, the obtained The probabilities of being a fake class {ω1} and a real class {ω0} are respectively Bayesian basic belief assignment BBA is: wherein, denotes the jth remote sensing test image in the test set; a soft classification result representing the test set; N te denotes the total number of test sets; representing a probability of forgery in the corresponding color space; denotes is the probability of being real in the corresponding color space; denotes a subset of the set Ω; Ω represents global uncertainty set; The weighted evidence fusion module fuses the evidence according to the optimal weight β n The soft classification result is discounted and simplified as: wherein, represents the BBA assigned to the A subset; denotes the BBA assigned to the global uncertainty set Ω; The rule of the weighted DS evidence fusion is: Fusion of the four groups of BBA results in a fusion result is: Wherein, K represents conflict factor; represents a DS rule-based evidence combination operator; B and C both represent subsets of set Ω; m1 represents BBA of information source 1; m2 represents BBA of information source 2; represents the fused result after discounting; The BetP probability form of the fused result is: Query image The authenticity decision is made by comparing the fused probability of forgery with the fused probability of authenticity to determine the authenticity discrimination result: The final authenticity identification result is output.

Images